Kokkinos, Filippos;
(2023)
Deep Structured Layers for Instance-Level Optimization in 2D and 3D Vision.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
Kokkinos_Main.pdf - Other Download (14MB) | Preview |
Abstract
The approach we present in this thesis is that of integrating optimization problems as layers in deep neural networks. Optimization-based modeling provides an additional set of tools enabling the design of powerful neural networks for a wide battery of computer vision tasks. This thesis shows formulations and experiments for vision tasks ranging from image reconstruction to 3D reconstruction. We first propose an unrolled optimization method with implicit regularization properties for reconstructing images from noisy camera readings. The method resembles an unrolled majorization minimization framework with convolutional neural networks acting as regularizers. We report state-of-the-art performance in image reconstruction on both noisy and noise-free evaluation setups across many datasets. We further focus on the task of monocular 3D reconstruction of articulated objects using video self-supervision. The proposed method uses a structured layer for accurate object deformation that controls a 3D surface by displacing a small number of learnable handles. While relying on a small set of training data per category for self-supervision, the method obtains state-of-the-art reconstruction accuracy with diverse shapes and viewpoints for multiple articulated objects. We finally address the shortcomings of the previous method that revolve around regressing the camera pose using multiple hypotheses. We propose a method that recovers a 3D shape from a 2D image by relying solely on 3D-2D correspondences regressed from a convolutional neural network. These correspondences are used in conjunction with an optimization problem to estimate per sample the camera pose and deformation. We quantitatively show the effectiveness of the proposed method on self-supervised 3D reconstruction on multiple categories without the need for multiple hypotheses.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Deep Structured Layers for Instance-Level Optimization in 2D and 3D Vision |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2023. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms. Access may initially be restricted at the author’s request. |
| UCL classification: | UCL UCL > Provost and Vice Provost Offices > UCL BEAMS UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science > Dept of Computer Science |
| URI: | https://discovery-pp.ucl.ac.uk/id/eprint/10164015 |
Archive Staff Only
 |
View Item |
